A rapid reading training control method based on eye tracking

Abstract

The present invention discloses a fast reading training control method based on eye tracking, including: extracting the fixation point coordinates and fixation point dwell time according to the eye movement trajectory curve, performing density clustering on the fixation points to obtain the sight distribution concentration area, calculating the center coordinates and average dwell time of each concentrated area, and generating a sight distribution heat map; calculating the reading efficiency score according to the reader's reading speed and the number of reviews, performing trend prediction on the reading efficiency score sequence, dynamically adjusting the font size according to the prediction result, and triggering font enlargement if the predicted reading efficiency shows a downward trend, and the font size adjustment range is determined by the reading efficiency decline rate. The present invention significantly improves the effect and experience of fast reading training through real-time monitoring, intelligent analysis and adaptive adjustment, and provides an innovative solution for improving personalized reading ability.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a fast reading training control method based on eye movement tracking. Background Art

[0002] Eye tracking decoding technology faces a complex technical challenge in practical applications. It is necessary to process eye movement data sampled at a high frequency in real time, which contains tiny eye movements and requires extremely high processing accuracy and speed. When dynamically adjusting the text presentation area, a balance must be struck between maintaining a smooth reading experience and accurately reflecting the cognitive rhythm. Too frequent adjustments may cause visual fatigue, while too slow a response will not achieve synchronization. More challenging is the need to accurately identify different types of eye movement patterns, such as saccades and dwells, and judge the reader's reading state based on this. This requires the algorithm to quickly analyze the speed, acceleration, and curvature characteristics of the eye movement trajectory, and use hidden Markov models to model the reading state in real time. At the same time, the font size needs to be dynamically adjusted according to the reading efficiency index, which requires trend prediction in a short time. However, the relationship between the reader's cognitive load state and eye movement characteristics is relatively complex. It is necessary to comprehensively consider multimodal data such as eye movement trajectory, pupil diameter, and blink frequency, and construct a cognitive load state space model through multi-view learning technology. How to effectively integrate these features and achieve joint adjustment of text presentation speed and font size on this basis constitutes a unique and challenging technical problem. This multi-dimensional real-time optimization problem places extremely high demands on algorithm design and scheduling of computing resources, and requires finding the best balance between improving reading efficiency and reducing cognitive burden. Summary of the invention

[0003] The purpose of the present invention is to solve the problems in the prior art and to propose a fast reading training control method based on eye movement tracking.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A fast reading training control method based on eye tracking, S1, using a high-frequency sampling eye tracking device to obtain reader eye movement data, performing noise elimination processing on the eye movement data, calculating an eye movement trajectory coordinate sequence based on the processed data, and performing smoothing processing on the coordinate sequence to obtain a continuous eye movement trajectory curve;

[0006] S2, extract the gaze point coordinates and the gaze point dwell time according to the eye movement trajectory curve, perform density clustering on the gaze points, obtain the gaze distribution concentration area, calculate the center coordinates and average dwell time of each concentration area, and generate the gaze distribution heat map;

[0007] S3, according to the sight distribution heat map, obtain the sight distribution concentrated area with the highest heat value, extract the boundary of the area, obtain the boundary polygon of the area of ​​interest, determine the zoom multiple and translation distance of the text presentation area according to the shape and size of the boundary polygon, and trigger the gradual zoom and translation of the text area;

[0008] S4, based on the eye movement data, extract and analyze the speed, acceleration and curvature characteristics of the eye movement trajectory to determine the reader's reading state, use the hidden Markov model to model the reading state, dynamically adjust the text presentation speed according to the transition probability matrix of the reading state, and reduce the adjustment amplitude of the text presentation speed if the reading state changes frequently, and calculate the reader's reading speed and number of reviews;

[0009] S5, calculates the reading efficiency score according to the reader's reading speed and the number of times of review, predicts the trend of the reading efficiency score sequence, and dynamically adjusts the font size according to the prediction result. If the predicted reading efficiency shows a downward trend, the font size is triggered to be enlarged;

[0010] S6 uses multi-view learning and combines multimodal data such as eye movement trajectory, pupil diameter and blinking frequency, as well as reading efficiency scores to construct a state-space model of readers' cognitive load. It then fuses the multimodal data through principal component analysis to obtain a comprehensive evaluation index of cognitive load. If the evaluation index exceeds the preset threshold, it triggers the joint adjustment of text presentation speed and font size.

[0011] Compared with the prior art, the present invention has the following advantages:

[0012] The present invention discloses a rapid reading training control method based on eye tracking. The method solves the problem of lack of personalization, real-time feedback and adaptive adjustment in traditional rapid reading training. The reader's eye movement data is obtained through a high-frequency eye tracking device, and accurate monitoring of the reading process is achieved. By using the sight distribution heat map and the area of ​​interest analysis, the intelligent zooming and translation of the text presentation area are realized, which effectively improves the reading efficiency. The reading state is modeled by a hidden Markov model, and the text presentation speed is dynamically adjusted to make the training process smoother and more personalized. Combined with the trend prediction of the reading efficiency score, the adaptive adjustment of the font size is achieved, which helps to relieve visual fatigue. The cognitive load state space model is constructed by multi-view learning technology, which realizes a comprehensive evaluation of the reader's cognitive state, and the text presentation speed and font size are jointly adjusted accordingly, effectively balancing the reading efficiency and cognitive load. In summary, this method significantly improves the effect and experience of rapid reading training through real-time monitoring, intelligent analysis and adaptive adjustment, and provides an innovative solution for improving personalized reading ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1A flowchart of a rapid reading training control method based on eye tracking proposed by the present invention;

[0014] Figure 2 This is a schematic diagram of a rapid reading training control method based on eye movement tracking proposed by the present invention. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0016] like Figure 1-Figure 2 The present embodiment provides a rapid reading training control method based on eye tracking, which specifically includes:

[0017] S1, use high-frequency sampling eye tracking equipment to obtain readers' eye movement data, including eye movement trajectory, pupil diameter and blinking frequency, perform noise elimination on the eye movement data, calculate the eye movement trajectory coordinate sequence based on the processed data, and smooth the coordinate sequence to obtain a continuous eye movement trajectory curve.

[0018] Specifically, the raw data of eye movement are obtained, and the raw data of eye movement are collected by a high-frequency sampling device; the nonlinear least square method is used to fit the raw data of eye movement, and abnormal points that deviate far from the fitting curve are eliminated, and the remaining data are subjected to noise elimination processing by a median filtering algorithm to obtain a preliminary denoised eye movement data sequence; based on the preliminary denoised eye movement data sequence, the coordinate sequence of the eye movement trajectory is calculated by a cubic spline interpolation method, and the coordinate sequence is filtered by applying a Gaussian kernel of the standard deviation to obtain a continuous and smooth eye movement trajectory curve.

[0019] Exemplarily, the feature vectors are normalized to eliminate the influence of different feature dimensions. The principal component analysis algorithm is used to reduce the dimension of the normalized feature vectors, and the principal components with an explained variance ratio of 95% are retained. Based on the feature vectors after dimension reduction, the K-means clustering algorithm is used to classify eye movement behaviors and obtain different types of eye movement patterns. In the eye tracking experiment, the original eye movement data with a sampling frequency of 100Hz were collected. The Levenberg-Marquardt algorithm was used for nonlinear least squares fitting, and the residual threshold was set to 2 standard deviations to remove abnormal points exceeding the threshold. The median filter of a 5×5 window was applied to the remaining data to obtain a preliminary denoised eye movement data sequence. Based on the denoised data, the cubic spline interpolation method was used to calculate the eye movement trajectory coordinates, and the interpolation point interval was set to 10ms. The interpolation coordinate sequence was filtered using a Gaussian kernel with a standard deviation of 1.5 pixels to obtain a continuous and smooth eye movement trajectory curve.

[0020] S2, extract the gaze point coordinates and gaze point dwell time according to the eye movement trajectory curve, perform density clustering on the gaze points to obtain the gaze distribution concentration area, calculate the center coordinates and average dwell time of each concentration area, and generate a gaze distribution heat map.

[0021] Specifically, the eye movement trajectory curve data is obtained, the fixation point coordinates and the corresponding dwell time are extracted from the eye movement trajectory curve data, and the eye movement speed is calculated according to the fixation point coordinates and the dwell time; the fixation point is determined by the adaptive threshold method, and the speed threshold is dynamically adjusted according to the local eye movement speed distribution. If the eye movement speed is lower than the speed threshold and the duration exceeds the preset time threshold, it is determined as a fixation point; DBSCAN density clustering is performed on the fixation point, and the DBSCAN density clustering sets the neighborhood radius as a first preset value and the minimum number of points as a second preset value to obtain the concentrated area of ​​sight distribution; for the concentrated area of ​​sight distribution, the dwell time is used The weighted average coordinate is calculated as the center coordinate, and the arithmetic mean of the dwell time of all gaze points in the gaze distribution concentration area is calculated as the average dwell time; a gaze distribution heat map is generated according to the gaze distribution concentration area, and the gaze distribution heat map adopts the Gaussian kernel density estimation method, and the standard deviation of the kernel function of the Gaussian kernel density estimation method is set to the third preset value, and the gaze point dwell time is used as the weight to calculate the thermal value of each pixel; the thermal value is normalized to obtain the normalized thermal value; RGB color coding is generated according to the normalized thermal value; the RGB color coding is superimposed on the original stimulus image to obtain the final gaze distribution heat map.

[0022] For example, in the eye tracking experiment, eye trajectory data with a sampling frequency of 100Hz was collected. The eye velocity sequence was obtained by calculating the Euclidean distance between adjacent sampling points and dividing it by the time interval of 10ms. The local eye velocity distribution was calculated using a sliding window of 20 sampling points, and the 15% quantile was taken as the dynamic velocity threshold. If the eye velocity of 10 consecutive sampling points is lower than the threshold, it is identified as a fixation point, and its average coordinates and 100ms dwell time are recorded. The DBSCAN algorithm was applied to the extracted fixation points, and the neighborhood radius ε=50 pixels and the minimum number of points MinPts=5 were set to obtain the concentrated area of ​​sight distribution. During the clustering process, points that were not divided into any clusters were marked as noise points and ignored in subsequent analysis. For each valid clustering area, the weighted average coordinates of the dwell time were calculated as the center coordinates, such as (Σtixi / Σti,Σtiyi / Σti), where ti is the dwell time of the i-th fixation point and (xi,yi) is its coordinate. At the same time, the arithmetic mean of the dwell time of all fixation points in the area was calculated as the average dwell time of the area. The heat map was generated using Gaussian kernel density estimation, with the kernel function standard deviation σ = 30 pixels. The fixation time was used as the weight to calculate the thermal value H(x,y) = Σ(ti·exp(-(x-xi)²-(y-yi)²) / (2σ²)) of each pixel at a resolution of 1920×1080. The obtained thermal values ​​were normalized to the minimum and maximum values, mapped to the color space of 0-255, and RGB color coding was generated. Finally, the RGB color coding was superimposed on the original stimulus image with a resolution of 1920×1080, and the transparency of the pixel with a thermal value of h was set to min(1,h·5) to achieve a visual display of the heat map.

[0023] S3, based on the gaze distribution heat map, obtain the gaze distribution concentrated area with the highest heat value, extract the boundary of the area, obtain the boundary polygon of the area of ​​interest, determine the zoom factor and translation distance of the text presentation area according to the shape and size of the boundary polygon, trigger the gradual zoom and translation of the text area, and the zoom rate and translation speed are determined by the average dwell time of the gaze point.

[0024] Specifically, according to the sight distribution heat map, the Otsu algorithm is used to adaptively determine the thermal value threshold, and the area above the thermal value threshold is binarized to obtain the candidate set area image; the Canny edge detection operator is applied to the candidate set area image to extract the boundary, and the low threshold and the high threshold are dynamically adjusted according to the image features to obtain the edge pixel point set; the edge pixel points are connected by the edge tracking algorithm to form a closed boundary polygon; the Douglas-Peucker algorithm is used to simplify the boundary polygon, and the tolerance of the Douglas-Peucker algorithm is set to a preset pixel value to obtain the simplified boundary polygon; the simplified The shape features and size parameters of the boundary polygon include the polygon area, perimeter, and aspect ratio of the minimum circumscribed rectangle; the zoom rate and translation speed are calculated based on the average dwell time of the gaze point, and the zoom rate and translation speed are obtained by mapping the average dwell time with an exponential decay function; if multiple boundary polygons are detected, they are sorted by area size, and a preset number of the largest polygons are selected as the main regions of interest for processing; the zoom factor of the text presentation area is calculated based on the polygon area and the image area, and the translation distance is determined based on the Euclidean distance from the center point of the polygon to the center of the image; for multiple regions of interest, the zoom factor and translation distance are weighted by area.

[0025] Exemplarily, the Douglas-Peucker algorithm is used to simplify the boundary polygon, and the tolerance is set to 2 pixels to reduce redundant vertices. If multiple boundary polygons are detected, they are sorted by area size, and the first three largest polygons are selected as the main regions of interest for processing. The shape characteristics and size parameters of the boundary polygons, including the polygon area, perimeter, and aspect ratio of the minimum circumscribed rectangle, are calculated, and the zoom factor and translation distance of the text presentation area are set, where the zoom factor = 1 + polygon area / image area, and the translation distance = the Euclidean distance from the center point of the polygon to the center of the image. The zoom rate and translation speed are calculated based on the average dwell time of the gaze point, and the exponential decay function is used to map the dwell time to the rate and speed parameters, the rate = the basic rate exp (-0.1 average dwell time), where the basic rate is preset to 1, triggering the gradual zoom and translation of the text area to achieve dynamic adjustment. For multiple regions of interest, the zoom and translation resources are weighted by area to ensure the coordination of the overall visual effect.

[0026] In a 1920x1080 resolution line of sight distribution heat map, the Otsu algorithm is used to adaptively calculate the heat value threshold, and the threshold is 0.6 (normalized heat value). The area with a heat value greater than 0.6 is binarized to obtain the candidate set area image. The Canny edge detection operator is applied to the binary image, and the low threshold is dynamically set to 50 and the high threshold is 150 through image gradient statistics to obtain the edge pixel set. The 8-connected edge tracking algorithm is used to connect the edge pixels to form three closed boundary polygons. The Douglas-Peucker algorithm is used to simplify each polygon, and the tolerance is set to 2 pixels, reducing the original number of vertices from an average of 150 to about 50. The three polygons are sorted by area size, accounting for 15%, 10% and 5% of the image area respectively. The shape features of the largest polygon are calculated: the area is 311040 pixels, the perimeter is 2234 pixels, and the aspect ratio of the minimum enclosing rectangle is 1.8. Based on this, the zoom factor of the text rendering area is set to 1.15, and the translation distance is 320 pixels, the Euclidean distance from the center of the image to the center of the polygon. Based on the average dwell time of the gaze point in this area of ​​250ms, the zoom rate is calculated to be 0.975 and the translation speed is 0.78 pixels / ms. For the second and third largest polygons, resources are allocated according to the area ratio, with zoom factors of 1.1 and 1.05 respectively, and translation distances of 240 pixels and 160 pixels. Finally, according to the calculation results, the gradual zoom and translation of the three text areas are triggered simultaneously to achieve coordinated dynamic visual effect adjustment.

[0027] S301, calculate the area and perimeter of the boundary polygon to obtain the shape complexity index, use the minimum circumscribed rectangle of the polygon to determine the main direction and aspect ratio, and establish a function model of the zoom factor and translation distance based on the shape complexity, main direction and aspect ratio. Consider the center of mass position of the polygon, adjust the translation direction and distance, set the maximum and minimum thresholds of zoom and translation, and ensure the rationality of text presentation.

[0028] Specifically, the area of ​​the boundary polygon is obtained by the polygon area calculation formula, and the distance between adjacent vertices of the boundary polygon is calculated by the Euclidean distance formula and the perimeter is accumulated. The shape complexity index is calculated by the logarithmic function of the area-to-perimeter ratio to determine the complexity of the boundary polygon. For multiple boundary polygons, sort them by area size, select the three largest ones for processing, and ignore the rest. The minimum bounding rectangle of the boundary polygon is calculated by the rotating caliper algorithm, the major axis direction of the minimum bounding rectangle is obtained as the main direction of the boundary polygon, the ratio of the major axis to the minor axis of the minimum bounding rectangle is calculated to obtain the aspect ratio, and the overall shape characteristics of the polygon are determined. If the aspect ratio is between 0.9 and 1.1, it is determined to be approximately circular, and the main direction is set to 0 degrees. According to the shape complexity index, the main direction and the aspect ratio, a function model of the scaling factor and the translation distance is established, and the third-order polynomial regression method is used to fit the functional relationship. The independent variables include shape complexity, main direction and aspect ratio, and the calculation formula of the scaling factor and the translation distance is obtained. The centroid coordinates of the boundary polygon are calculated, and the translation direction and distance are adjusted in combination with the centroid coordinates and the main direction. Calculate the vector from the centroid coordinates to the image center and use the vector as the reference direction for translation. Set the upper and lower thresholds of the zoom factor to 0.5 and 2, and the upper and lower thresholds of the translation distance to 10% and 50% of the image width to ensure the rationality of text presentation. Use the output of the function model for subsequent dynamic adjustments to achieve adaptive changes in the text area.

[0029] For example, in an image with a resolution of 1920x1080, 5 boundary polygons are detected. First, the area and perimeter of each polygon are calculated. The largest polygon has an area of ​​120,000 pixels and a perimeter of 1,500 pixels. The shape complexity index is calculated by ln(4π*area / perimeter²) and is 0.85. After sorting by area size, the first three polygons are selected for processing. The rotating caliper algorithm is applied to the largest polygon to obtain the minimum enclosing rectangle with a major axis direction of 30 degrees and an aspect ratio of 1.8. Since the aspect ratio is not between 0.9 and 1.1, 30 degrees is retained as the main direction. Using a third-order polynomial regression, with a shape complexity of 0.85, a main direction of 30 degrees, and an aspect ratio of 1.8 as input, a scaling factor of 1.3 and a translation distance of 200 pixels are calculated. The coordinates of the center of mass of the polygon are calculated to be (960,540), and the vector (0,0) from the center of mass to the center of the image (960,540) is obtained. Since the vector is zero, the main direction of 30 degrees is used as the translation direction. The scale factor of 1.3 is limited to between 0.5 and 2, and the translation distance of 200 pixels is limited to between 192 and 960 pixels (10% to 50% of the image width). Repeat the above process for the second and third largest polygons to obtain their respective scale factors and translation distances.

[0030] S4, based on the eye movement data, extracts and analyzes the speed, acceleration and curvature characteristics of the eye movement trajectory to determine the reader's reading state, uses the hidden Markov model to model the reading state, and dynamically adjusts the text presentation speed according to the transition probability matrix of the reading state. If the reading state changes frequently, the text presentation speed adjustment amplitude is reduced, and the reader's reading speed and number of reviews are calculated.

[0031] Specifically, based on the eye movement data, median filtering is used to remove outliers, and the window size of the median filter is set to a set number of sampling points; the speed sequence is obtained by calculating the distance between adjacent sampling points and dividing it by the time interval, and the acceleration sequence is obtained by calculating the difference of the speed sequence. The curvature sequence is calculated using the three-point method, and the speed, acceleration and curvature characteristics of the eye movement trajectory are obtained by sliding window smoothing with set sampling points; the thresholds of speed, acceleration and curvature are adaptively adjusted according to the difficulty of the text, and the eye movement characteristics are divided into three basic states: fixation, saccade and return; the hidden Markov model is used to model the reading state, and the model parameters are trained through the Baum-Welch algorithm to obtain the state transition probability. Matrix and observation probability distribution; use the model parameters obtained by training, apply the Viterbi algorithm to decode the eye movement feature sequence, and obtain the most likely reading state sequence; calculate the state transition frequency, if the state transition frequency exceeds the preset threshold, reduce the text presentation speed adjustment amplitude; for multiple paragraphs of text, count the number of fixations and duration in each reading state sequence, calculate the average fixation time, calculate the reading speed in combination with the total number of words in the text, and count the number of occurrences of the look-back state to obtain the number of reviews; take the average value of all paragraphs as the quantitative feature of the overall reading behavior; dynamically adjust the text font size, line spacing and paragraph spacing according to the current reading state and reading speed to achieve adaptive typesetting;

[0032] For example, in a reading experiment, eye movement data with a sampling frequency of 100Hz were collected. First, a 5-point median filter was used to remove outliers, and then the speed, acceleration and curvature sequences were calculated. A 20-point sliding window smoothing process was applied to these sequences to obtain smoothed feature sequences. According to the text difficulty coefficient of 0.7 (in the range of 0-1), the speed threshold was dynamically set to 20° / s, the acceleration threshold was 400° / s², and the curvature threshold was 0.5. These thresholds were used to divide the eye movement features into three states: fixation, saccade and return. A 3-state hidden Markov model was used, and the Baum-Welch algorithm was trained for 100 iterations to obtain the state transition probability matrix and observation probability distribution. Using the trained model, the Viterbi algorithm was used to decode the eye movement feature sequence of 1 minute to obtain the most likely reading state sequence. The state transition frequency was calculated to be 2.5 times / second, which exceeded the preset threshold of 2 times / second, so the text presentation speed adjustment was reduced from the original 10% to 5%. The text presentation speed was dynamically adjusted using the exponential smoothing method (α=0.2) from the initial value of 120 words / minute to 110 words / minute. For reading materials containing three paragraphs of text, the average fixation time (250ms, 220ms, 280ms), reading speed (180 words / minute, 200 words / minute, 160 words / minute) and number of reviews (5 times, 3 times, 7 times) of each paragraph were calculated, and the average value was taken to obtain the overall reading behavior characteristics. Finally, based on the current slow reading state, the text font size was increased from 16pt to 18pt, the line spacing was increased from 1.2 times to 1.4 times, and the paragraph spacing was increased from 1.5 times the line height to 2 times the line height.

[0033] S5 calculates the reading efficiency score based on the reader's reading speed and number of review times, predicts the trend of the reading efficiency score sequence, and dynamically adjusts the font size based on the prediction results. If the predicted reading efficiency shows a downward trend, font enlargement is triggered, and the font size adjustment range is determined by the reading efficiency decline rate.

[0034] Specifically, according to the obtained reading speed and the number of reviews, the reading speed and the number of reviews are z-score standardized, and the reading efficiency score is calculated by the weighted summation method; the reading efficiency score is stored in a circular data queue of a preset length, and the short-term trend and the long-term trend are calculated for the reading efficiency score sequence in the circular data queue; the reading efficiency is predicted by the exponential smoothing method, and the smoothing coefficient in the exponential smoothing method is set to a predetermined value, and the prediction result is compared with the actual value at the previous moment to determine the trend of reading efficiency changes; if it is determined that the reading efficiency is on a downward trend, the font enlargement mechanism is triggered, the reading efficiency decrease rate is calculated, and the decrease rate is mapped to the font size adjustment range by the sigmoid function; if it is determined that the reading efficiency is on an upward trend, the font reduction mechanism is triggered, and the font size adjustment range is calculated; the font size is updated according to the font size adjustment range, and the updated font size is subject to preset upper and lower thresholds; the updated font size is smoothed by the moving average method to obtain the smoothed font size, and the smoothed font size is recorded as the reference value for the next adjustment.

[0035] Exemplarily, according to the reader's reading speed and the number of reviews, the two are first z-score standardized, and then the weighted summation method is used to calculate the reading efficiency score, the reading speed weight is set to 0.7, the review number weight is set to 0.3, and the calculated reading efficiency score sequence is stored in a circular data queue with a length of 100. The short-term (5 minutes) and long-term (30 minutes) trends of the reading efficiency score sequence are calculated simultaneously, and the exponential smoothing method is used for prediction. The smoothing coefficient α is set to 0.2, the predicted value is calculated, and the predicted result is compared with the actual value of the previous moment to determine the changing trend of the reading efficiency. If the predicted reading efficiency shows a downward trend, the font enlargement mechanism is triggered, the reading efficiency decline rate is calculated, and the sigmoid function is used to map the decline rate to the font size adjustment range. The function parameters are set so that when the decline rate is in the range of 0 to 0.5, the font size increases by 1 to 5 pixels. When the predicted reading efficiency shows an upward trend, the font size is gradually reduced, but the rate of change is set to 1 / 3 of the downward trend. The upper and lower thresholds of font size adjustment are set, with the minimum font size of 12 pixels and the maximum font size of 24 pixels. The font size is updated according to the calculated adjustment range. The moving average method is used to smooth the change of font size. The window size is set to 3 adjustments, and the adjusted font size is recorded as the benchmark value for the next adjustment. In a reading experiment, the reader's reading speed was recorded as 180 words / minute and the number of reviews was 5 times / page. First, the two indicators were z-score standardized, and the standardized reading speed was 0.8 and the number of reviews was -0.5. The weighted sum of weights 0.7 and 0.3 was used to calculate the reading efficiency score of 0.41. This score was stored in a circular data queue with a length of 100. At the same time, the reading efficiency trends of 5 minutes and 30 minutes were calculated, and the exponential smoothing method (α=0.2) was used for prediction. The 5-minute trend prediction value is 0.39, and the 30-minute trend prediction value is 0.43. Compared with the actual value of 0.42 at the previous moment, it shows a downward trend in the short term. The font enlargement mechanism is triggered, and the reading efficiency decrease rate is calculated to be 0.07. The sigmoid function is used to map this decrease rate to the font size adjustment amplitude, and the increase value is 2.1 pixels. The current font size is 16 pixels, and after adding 2.1 pixels, it is 18.1 pixels. The moving average method is used for smoothing, and the 18 pixels and 17 pixels of the previous two adjustments are considered. Finally, the new font size is determined to be 17.7 pixels. This value is recorded as the benchmark for the next adjustment, and the actual displayed font size is set to 18 pixels to ensure that it is within the range of 12 to 24 pixels.

[0036] S6 uses multi-view learning technology, combines multimodal data such as eye movement trajectory, pupil diameter and blinking frequency, and reading efficiency score to construct a reader's cognitive load state space model. The multimodal data is fused through principal component analysis to obtain a comprehensive evaluation index of cognitive load. If the evaluation index exceeds the preset threshold, the joint adjustment of text presentation speed and font size is triggered.

[0037] Collect multimodal data of eye movement trajectory, pupil diameter and blinking frequency, and the multimodal data is obtained by eye tracking devices with different sampling frequencies; execute the cubic spline interpolation algorithm based on the multimodal data to obtain a synchronized eye movement data sequence; perform feature extraction and preprocessing on the synchronized eye movement data sequence, and construct a multidimensional feature vector containing the eye movement trajectory speed, pupil diameter change rate and blinking frequency time series; use the multi-view learning algorithm to perform feature learning and fusion on the multidimensional feature vector, use the consensus principal component analysis method to map the features of different views to a common latent space, use the adaptive weight method to dynamically adjust the weight according to the discrimination of each feature, and obtain a unified feature representation; based on the unified feature representation, construct a reader's cognitive load state space model, use the Kalman filter algorithm to estimate and predict the state, set the state transfer matrix and observation matrix, initialize the state covariance matrix, and obtain a real-time cognitive load state estimate; set the threshold of the cognitive load evaluation indicator, and if the cognitive load state estimate exceeds the preset threshold, trigger the joint adjustment mechanism of text presentation speed and font size, and adjust the text presentation speed and font size proportionally. According to the degree of exceeding the threshold, the text rendering speed and font size are adjusted proportionally, with the text rendering speed reduced by 5 words per second and the font size increased by 1 pixel each time to achieve dynamic adaptation. The adjustment effect is evaluated through short-term changes in reading efficiency, forming a closed-loop feedback mechanism to continuously optimize the adjustment strategy.

[0038] For example, in a reading experiment, 100Hz eye movement trajectory, 50Hz pupil diameter and 25Hz blink frequency data were collected simultaneously. All data were synchronized to 100Hz using cubic spline interpolation. The speed and acceleration of the eye movement trajectory were calculated, and the average speed was 200 pixels / second and the average acceleration was 500 pixels / second²; the change rate of pupil diameter was calculated, with an average of 0.2mm / second; the blink frequency was statistically 20 times / minute. Combined with the reading efficiency score of 0.75, an 8-dimensional feature vector was constructed. The consensus principal component analysis method was used for feature fusion, and the initial value of the feature weight was set to 0.125. The feature discrimination was iteratively calculated and the weight was adjusted to [0.15, 0.18, 0.12, 0.13, 0.11, 0.10, 0.11, 0.10]. The Kalman filter algorithm was used to construct a cognitive load state space model. The state transfer matrix was set to the unit matrix, the observation matrix was [1,1,1,1,1,1,1,1], and the initial state covariance matrix was the unit matrix multiplied by 0.1. After filtering, the estimated value of the cognitive load state was 0.82, which exceeded the preset threshold of 0.8. The adjustment mechanism was triggered to reduce the text presentation speed from the original 300 words / minute to 295 words / minute, and at the same time increase the font size from 16 pixels to 17 pixels. In the next minute, it was monitored that the short-term reading efficiency increased by 5%, verifying the effectiveness of the adjustment.

[0039] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A rapid reading training control method based on eye tracking, characterized in that: The following steps are involved: S1, using a high-frequency sampling eye tracking device to obtain the reader's eye movement data, performing noise elimination on the eye movement data, calculating the eye movement trajectory coordinate sequence based on the processed data, and smoothing the coordinate sequence to obtain a continuous eye movement trajectory curve; S2, extract the gaze point coordinates and the gaze point dwell time according to the eye movement trajectory curve, perform density clustering on the gaze points, obtain the gaze distribution concentration area, calculate the center coordinates and average dwell time of each concentration area, and generate the gaze distribution heat map; S3, according to the sight distribution heat map, obtain the sight distribution concentrated area with the highest heat value, extract the boundary of the area, obtain the boundary polygon of the area of ​​interest, determine the zoom multiple and translation distance of the text presentation area according to the shape and size of the boundary polygon, and trigger the gradual zoom and translation of the text area; S4, based on the eye movement data, extract and analyze the speed, acceleration and curvature characteristics of the eye movement trajectory to determine the reader's reading state, use the hidden Markov model to model the reading state, dynamically adjust the text presentation speed according to the transition probability matrix of the reading state, and reduce the adjustment amplitude of the text presentation speed if the reading state changes frequently, and calculate the reader's reading speed and number of reviews; S5, calculates the reading efficiency score according to the reader's reading speed and the number of times of review, predicts the trend of the reading efficiency score sequence, and dynamically adjusts the font size according to the prediction result. If the predicted reading efficiency shows a downward trend, the font size is triggered to be enlarged; S6 uses multi-view learning and combines multimodal data such as eye movement trajectory, pupil diameter and blinking frequency, as well as reading efficiency scores to construct a state-space model of readers' cognitive load. It then fuses the multimodal data through principal component analysis to obtain a comprehensive evaluation index of cognitive load. If the evaluation index exceeds the preset threshold, it triggers the joint adjustment of text presentation speed and font size.

2. The method for rapid reading training control based on eye tracking according to claim 1, characterized in that: The S1 includes: Obtaining raw eye movement data, including eye movement trajectory, pupil diameter and blink frequency. The raw eye movement data is collected by high-frequency sampling equipment; The nonlinear least square method is used to fit the original eye movement data, and the abnormal points that deviate far from the fitting curve are eliminated. Then, the remaining data is subjected to noise elimination to obtain a preliminary denoised eye movement data sequence. Based on the preliminary denoised eye movement data sequence, the coordinate sequence of the eye movement trajectory is calculated using the cubic spline interpolation method. The coordinate sequence is filtered using a Gaussian kernel with standard deviation to obtain a continuous and smooth eye movement trajectory curve.

3. The rapid reading training control method based on eye tracking according to claim 1, characterized in that: The S2 comprises: Obtaining eye movement trajectory curve data, extracting gaze point coordinates and corresponding dwell time from the eye movement trajectory curve data, and calculating eye movement velocity according to the gaze point coordinates and dwell time; Adopting the adaptive threshold method to determine the fixation point, the adaptive threshold method dynamically adjusts the speed threshold according to the local eye movement speed distribution. If the eye movement speed is lower than the speed threshold and the duration exceeds the preset time threshold, it is determined as a fixation point. Perform DBSCAN density clustering on the gaze points, set the neighborhood radius to the first preset value, and the minimum number of points to the second preset value, to obtain the concentrated area of ​​sight distribution; For the area where the sight distribution is concentrated, the dwell time is used as the weight to calculate the weighted average coordinate as the center coordinate, and the arithmetic mean of the dwell time of all the gaze points in the area where the sight distribution is concentrated is calculated as the average dwell time; Generate a sight distribution heat map based on the sight distribution concentration area.

4. The method for rapid reading training control based on eye tracking according to claim 1, characterized in that: The S3 includes: According to the sight distribution heat map, the Otsu algorithm is used to adaptively determine the heat value threshold, and the area above the heat value threshold is binarized to obtain the candidate concentrated area image; Apply the Canny edge detection operator to extract the boundary of the regional image in the candidate set, dynamically adjust the low threshold and high threshold according to the image features, and obtain the edge pixel point set; The edge pixels are connected by edge tracking algorithm to form a closed boundary polygon; The boundary polygon is simplified using the Douglas-Peucker algorithm, and the tolerance of the Douglas-Peucker algorithm is set to a preset pixel value to obtain a simplified boundary polygon; Calculate the shape characteristics and size parameters of the simplified boundary polygon, the shape characteristics and size parameters including the polygon area, perimeter and aspect ratio of the minimum circumscribed rectangle; The zoom rate and translation speed are calculated based on the average dwell time of the fixation point. The zoom rate and translation speed are obtained by mapping the average dwell time with an exponential decay function. If multiple boundary polygons are detected, they are sorted by area size and the largest polygons of a preset number are selected as the main regions of interest for processing; Calculate the scaling factor of the text presentation area according to the polygon area and the image area, and determine the translation distance according to the Euclidean distance from the center point of the polygon to the center of the image; For multiple regions of interest, zoom factors and translation distances are allocated by area weighting.

5. The method for rapid reading training control based on eye tracking according to claim 4, characterized in that: The S3 further includes: S301, calculate the area and perimeter of the boundary polygon to obtain the shape complexity index, use the minimum circumscribed rectangle of the polygon to determine the main direction and aspect ratio, and establish a function model of the zoom factor and translation distance based on the shape complexity, main direction and aspect ratio. Consider the center of mass position of the polygon, adjust the translation direction and distance, set the maximum and minimum thresholds of zoom and translation, and ensure the rationality of text presentation.

6. The method for rapid reading training control based on eye tracking according to claim 5, characterized in that: The S301 includes: The polygon area calculation formula is used to obtain the area of ​​the boundary polygon, and the Euclidean distance formula is used to calculate the distance between adjacent vertices of the boundary polygon and accumulate them to obtain the perimeter; The shape complexity index is calculated through the logarithmic function of the area-to-perimeter ratio to determine the complexity of the boundary polygon; The minimum enclosing rectangle of the boundary polygon is calculated by using the rotating caliper algorithm, the major axis direction of the minimum enclosing rectangle is obtained as the main direction of the boundary polygon, and the ratio of the major axis to the minor axis of the minimum enclosing rectangle is calculated to obtain the aspect ratio; According to the shape complexity index, main direction and aspect ratio, a function model of scaling factor and translation distance was established, and the third-order polynomial regression method was used to fit the functional relationship. Calculate the centroid coordinates of the boundary polygon, and adjust the translation direction and distance based on the centroid coordinates and the main direction; Calculate the vector from the centroid coordinate to the image center and use the vector as the reference direction of translation; The output of the function model is used for subsequent dynamic adjustments to adaptively change the text area.

7. The method for rapid reading training control based on eye tracking according to claim 1, characterized in that: The S4 comprises: Based on the eye movement data, median filtering was used to remove outliers, and the window size of the median filter was set to a set number of sampling points; The velocity sequence is obtained by calculating the distance between adjacent sampling points and dividing it by the time interval. The acceleration sequence is obtained by calculating the difference of the velocity sequence. The curvature sequence is calculated using the three-point method. The velocity, acceleration and curvature characteristics of the eye movement trajectory are obtained by using a sliding window smoothing process with set sampling points. The thresholds of speed, acceleration and curvature are adaptively adjusted according to the text difficulty, and the eye movement characteristics are divided into three basic states: fixation, saccade and return. The hidden Markov model is used to model the reading state, and the model parameters are trained by the Baum-Welch algorithm to obtain the state transition probability matrix and observation probability distribution; Using the trained model parameters, the Viterbi algorithm is used to decode the eye movement feature sequence to obtain the most likely reading state sequence. Calculate the state transition frequency, and if the state transition frequency exceeds a preset threshold, reduce the text presentation speed adjustment amplitude; For multiple paragraphs of text, the number and duration of fixations in each reading state sequence were counted, the average fixation time was calculated, and the reading speed was calculated based on the total number of words in the text. The number of occurrences of the look-back state was counted to obtain the number of reviews. The average value of all paragraphs was taken as a quantitative feature of the overall reading behavior; Dynamically adjust text font size, line spacing and paragraph spacing according to the current reading status and reading speed, and achieve adaptive typesetting.

8. The method for rapid reading training control based on eye tracking according to claim 1, characterized in that: The S5 includes: According to the obtained reading speed and review times, the reading speed and review times were z-score standardized, and the reading efficiency score was calculated using the weighted summation method; The reading efficiency scores are stored in a circular data queue of a preset length, and the short-term trend and the long-term trend are calculated for the reading efficiency score sequence in the circular data queue; The exponential smoothing method is used to predict reading efficiency. The smoothing coefficient in the exponential smoothing method is set to a predetermined value. The prediction result is compared with the actual value at the previous moment to determine the trend of reading efficiency change. If it is determined that the reading efficiency is on a downward trend, the font enlargement mechanism is triggered, the reading efficiency decrease rate is calculated, and the sigmoid function is used to map the decrease rate to the font size adjustment range; If it is determined that the reading efficiency is on an upward trend, the font reduction mechanism is triggered and the font size adjustment range is calculated; The font size is updated according to the font size adjustment amplitude, and the updated font size is subject to preset upper and lower limit thresholds; The updated font size is smoothed using a moving average method to obtain a smoothed font size, and the smoothed font size is recorded as a reference value for the next adjustment.

9. The method for rapid reading training control based on eye tracking according to claim 1, characterized in that: The S6 comprises: Collect multimodal data of eye movement trajectory, pupil diameter and blink frequency. The multimodal data is obtained by eye tracking devices with different sampling frequencies. A cubic spline interpolation algorithm is executed according to the multimodal data to obtain a synchronized eye movement data sequence; Feature extraction and preprocessing of synchronized eye movement data sequences were performed to construct a multi-dimensional feature vector containing the time series of eye movement trajectory velocity, pupil diameter change rate, and blink frequency; The multi-view learning algorithm is used to learn and fuse the multi-dimensional feature vectors, and the consensus principal component analysis method is used to map the features of different views into a common latent space to obtain a unified feature representation; Based on the unified feature representation, the reader cognitive load state space model is constructed, and the Kalman filter algorithm is used to estimate and predict the state to obtain the real-time cognitive load state estimation value; If the estimated value of the cognitive load state exceeds a preset threshold, the joint adjustment mechanism of the text presentation speed and the font size is triggered to adjust the text presentation speed and the font size proportionally.

Images